Context
Liu and colleagues introduce Agentic-R, a retriever training framework that combines intermediate-query passage relevance with final-answer correctness and alternates retriever and search-agent optimization. Controlled tests span seven question-answering benchmarks and three agent architectures.
What changed
A January 2026 original research preprint demonstrates agent-specific retriever training on intermediate search queries and final-answer utility, including a two-iteration agent–retriever optimization loop. Its authors report higher average exact-match scores than the evaluated baselines.
Why it matters
In multi-step AI discovery, passages selected for semantic similarity to a subquery may not help complete the final answer. Retriever training objectives can therefore change which evidence reaches an agent, independently of traditional search ranking or publisher content changes.
Visibility OS interpretation
Visibility OS interpretation: distinguish query generation, local passage relevance, downstream evidence utility, and answer outcome when measuring agentic source selection. This is research on experimental retrieval systems, not evidence that commercial AI search products use Agentic-R or that publishers gain citations by adopting any specific page format.
What remains uncertain
Do the results replicate in live web-scale search, shopping and expert tasks? How well do the method and costs scale to larger agents and retrievers? Would citation faithfulness or publisher-level exposure change? The authors report no additional benefit from a third optimization iteration.
Evidence claims
The Agentic-R paper proposes training a search retriever using both LLM-scored relevance of candidate passages to intermediate agent queries and final-answer correctness, followed by alternating agent and retriever optimization.
In the authors' seven-benchmark evaluation with their trained search agent, Agentic-R achieved a 45.00 average exact-match score, compared with 41.78 for the strongest evaluated baseline REPLUG.
In the authors' Table 1 cross-agent evaluations, Agentic-R achieved average exact-match scores of 43.64 on R1-Searcher and 39.43 on SimpleDeepSearcher, exceeding the best listed comparison scores of 41.67 and 37.69 respectively.
Related historical records
LRAT trains search retrievers on agent browsing and reasoning trajectories
This record related to the linked event.